Cross-View Action Recognition via Transferable Dictionary Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Action recognition models often struggle with generalizing across different camera views.
• Discriminative features effective in fixed views may not transfer well to new perspectives.

Purpose of the Study:

• To develop robust methods for action recognition across varying camera views.
• To enhance the generalization capability of action recognition models to unseen views.

Main Methods:

• Proposed two dictionary learning approaches: view-specific dictionaries and a common dictionary for shared features.
• Learned dictionaries simultaneously from videos across different views to encourage similar sparse representations.
• Utilized both view-specific and common dictionaries for representing videos from each view.

Main Results:

• The common dictionary enables representation of actions from unseen views.
• The approach demonstrated effectiveness in semi-supervised learning scenarios with limited labeled data.
• Achieved superior performance compared to existing methods on three public datasets for cross-view action recognition.

Conclusions:

• The proposed dictionary learning framework significantly improves cross-view action recognition.
• The common dictionary is key to handling novel viewpoints and semi-supervised learning.
• This work offers a promising direction for more adaptable and generalizable action recognition systems.